G-2.0-Code — Agentic Coding

G-2.0-Code is the second generation of our agentic coding model — purpose-built for the problems that matter in real software engineering: tracing bugs across large codebases, applying multi-step refactors without breaking the build, and calling APIs correctly in agentic loops.

A defining strength is function calling. The Berkeley Function Calling Leaderboard v4 measures how precisely a model invokes the right function with the right arguments across diverse APIs and schema formats. G-2.0-Code scores 72 on BFCL-V4. For production agentic systems, this translates directly to fewer failed tool calls, less error recovery, and more tasks completed without human intervention.

SWE-bench Verified at 77 percent places G-2.0-Code among capable coding models for real-world software engineering tasks. The model receives real GitHub issues from open-source projects and must produce patches that pass the test suite — no scaffolding, no hints.

The sparse Mixture-of-Experts architecture delivers strong code quality at efficient inference cost, and context expands to long context for repository-scale understanding. G-2.0-Code can hold entire monorepos in a single pass — understanding the full project structure before making any change.

Your code is protected by Guardian's privacy-first architecture — encrypted in transit and at rest, with minimal retention by default. Available now through the Onora App.